European Institute of Oncology

The first Global Dialogue on AI Governance needs to be successful and provide real and actionable results that transform principles into coordinated global implementation in high-impact sectors such as health care. The first is that it ought to establish a harmonised set of trustworthy AI guiding principles with particular relevance to clinical settings. More about how AI safety, transparency and accountability in healthcare is not negotiable since AI systems directly affect patient outcomes We would also need a globally aligned framework specific to high-risk applications that would be a key milestone. The Dialogue should facilitate closer alignment of regulatory pathways of clinical AI. Fragmented processes for approval and validation of new technologies today hinder the pace of innovation and create uncertainty for clinicians and patients. Also, achieving a secure and fair data collaboration should be the ultimate goal. The objective would involve creating multistakeholder task forces on key topics (eg, health care), getting together policy makers, clinicians, researchers, and industry. Those groups should be required to translate principles into things like operational guidelines and pilot programs. The Dialogue should ultimately outline specific and measurable ways forward—name what will happen, with when, pilot projects that may ensue, and an accountability mechanism. In the end, success will be evaluated on moving from dialogue to action in ways to enable AI to enhance patient outcomes with safety, equity, and global alignment.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight
• Interoperability of governance approaches

These priorities complement each other in highlighting the need to conduct AI safely, equitably and with clinical responsibility in high-impact domains including health. Together, these three priorities highlight the need for AI to be conducted with safety, equity, and clinical responsibility in high-impact areas such as health care .Safety, security, reliability and trustworthiness of AI must be maintained when it directly affects clinical decisions and thereby clinical results. For example when an AI model accurately predicts treatment choice or survival time, inaccuracy cannot be tolerated.Transparency, accountability and human oversight are interrelated goals. Ultimately, AI ought to assist in but not replace medical decisions by doctors practicing medicine in a clinical setting. In order to maintain trust and ensure patient safety, we believe that interpretability of systems and continued responsibility of clinicians cannot be neglected. In addition, it should also ensure interoperability of governance initiatives. This can lead to fragmented regulatory frameworks that do not allow for the exchange or use of AI tools that have shown to provide value in one region or market. This would support the safe and effective integration of AI into health systems at both country and international levels under aligned standards. Finally, we must attend to the social, economic, ethical, and technical aspects of AI to ensure we can avoid negative externalities. This risk include in healthcare an example of which is, datasets used to develop algorithms may not be representative of the population, thereby reinforcing or worsening existing health disparities. Hence the need for data practices that are inclusionary and globally representative. Combined, these priorities underpin a governance model around responsible innovation that must take place to ensure that AI people meaningful and fair benefits to patients and healthcare systems across the world.

Although the major themes identified reflect a range of important dimensions of AI governance, a number of cross-cutting and emerging issues warrant greater focus. Firstly, the ongoing challenge of clinical integration and workflow alignment is still less examined. It can also be with tools to truly test and make sure your AI systems can be meaningful in real life environments, like healthcare systems, with minimal disruption of existing workflows and burden on professionals. Instead, governance frameworks need to deal with improvement not just performance, but also usability, integration, and effect on decision-making processes. A second emerging gap is how to evaluate and govern life cycle processes and standards. It is important to note that AI systems can change (adaptive nature or machine learning). It needs governance that go beyond initial validation to include continuous monitoring and re-validation as well as post-deployment auditing. Third, the topic of data ownership is still poorly articulated. Issues about who controls data and how it is shared is still present. We need distinct governance models that can provide fairness, trust and sustainability. Fourth, and not the least important, the lack of education and AI literacy of professionals themselves. Especially in fields like healthcare, the safe and effective deployment of AI relies upon users having a good understanding of what it can, and cannot, do. This is why capacity-building should be viewed as a pillar of governance. Last but not least, the increasing use of AI in making critical decisions, especially in situations where uncertainty is unavoidable, faces the dilemma of who is liable for the errors, as well as howmuch risk human beings are willing to take when the consequences affect human beings directly. We have seen that given the nexus-like qualities of many of these cross-cutting issues, studying them together will be critical for ensuring AI governance frameworks are fit-for-purpose, context-aware, and responsive to implementation challenges in practice.

In the healthcare space, the governance gaps in AI is already having clinical and innovation impacts. Fragmentation of policy frameworks in different regions is one of the major challenges. While evolving regulation in Europe provides essential guardrails, divergent thoroughness and interpretation of those guardrails will slow clinical implementation of validated AI tools. Such uncertainty may delay translation from development to the clinic, hampering the ability of clinicians and researchers to apply research to real-world use. The second and related area of challenge is clinical validation/evidence generation. Most AI models demonstrate good performance from retrospective studies but do not have prospective validation and real-world evidence. In the absence of aligned standards for evaluation, it is impossible to know which tools can safely and effectively be embedded in routine clinical practice. Access to and representativeness of data continue to be major challenges as well. While all this should take place within strict and protective data protection frameworks, this strictness also makes it impossible to share data and combine datasets across institutional borders as we can be more certain that the data we work on will not contain sensitive information and that we have large data sets. It can result in nothing but non-generalizability of the models and the possibility of amplifying the current inequities in the healthcare end results. However, these gaps also offer meaningful opportunities. Harmonized regulatory pathways are gaining traction, allowing the riskier parts of AI to be deployed faster and more safely acrpss Health Systems. Efforts that promote secure, privacy-preserving data sharing may facilitate collaboration, as well as increase robustness of models. Moreover, rising awareness about the necessity of transparent and human-centered AI is driving the creation of more interpretable and clinically useful tools. This is especially true in oncology, where artificial intelligence has the ability to enhance the diagnosis, caret up, and outcome prediction. Figuring out how to help close these governance gaps is vital to marrying innovation with patient safety so that AI can yield impactful and equitable value in clinical care.

The AI Dialogue can serve as such a neutral, multistakeholder platform, providing a space and process to cross-cut geographical, sectoral, and disciplinary divides in pursuit of coherent, inclusive international AI governance. In the first place, it may assist in achieving harmonisation of values and norms in countries. Numerous AI governance frameworks are already being proposed and promoted, but the fragmentation of these frameworks stifles interoperability and hamstrings uptake, especially important in high impact areas like healthcare. The Dialogue can aid the convergence on common definitions, types of risk levels, and minimum thresholds of safety, transparency, and accountability. Second, Dialogue can facilitate knowledge transfer and best practice dissemination. It can facilitate cross-sector learning and expedite the horizontal transfer of effective governance models in fields developing rapidly such as clinical AI by gathering together policymakers, researchers, clinicians, industry leaders and civil society. Third, it may be a catalyst for joint actions and pilots. Beyond mere discussion, the Dialogue could assist in establishing international working groups aimed at producing tangible results – be it in the form of mechanisms for cross-border data exchange, assessment criteria, or regulatory sandboxes. Fourth, the Dialogue could to help promote inclusion and representation, ensuring that the people of less developed regions and under-represented groups have a voice in shaping AI governance.It is crucial to avoid possible global inequalities and to ensure that everyone has an equal share in AI's benefits.Finally, the Dialogue will promote accountability and continuity through follow up arrangements, progress monitoring and continuity between all too infrequent opportunities for global convergences.It is hoped that the AI Dialogue can help to achieve the dual objective of effective AI governance—for which global consensus and concerted action are necessary—thus grounding it in practice enough people will take notice to prevent errors.

The AI Dialogue must be a rallying point for existing global, regional and national initiatives, not just to avoid duplication, but to catalyse progress in a coherent manner toward overall AI governance. On the international front, work in organizations such as UNESCO and the OECD and process ongoing within the G7 and G20 have put in place high-level principles for trust and human-centered AI. Simultaneously, more tailored industry-specific initiatives, especially in health care, are establishing thresholds for clinical evidence, safety, and ethics of AI to meet the needs of operationalizing AI in the real world. At the regional level, the EU has stepped up in a big way with the creation of all-encompassing regulatory regimes that intend to balance innovation and a risk-based framework for oversight over it. Such efforts are a step towards establishing an organized framework that may guide large-scale concerted global harmonization, particularly with high-risk clinical AI applications. But national strategies or platforms, the Italian national AI initiatives (for example, ENIA), should be developed to turn principles into actions. Most of these initiatives are aimed at context-specific issues, focusing on data governance, infrastructure, and capacity-building, and have the potential to become effective testing grounds for scalable and clinically feasible solutions. The value of the AI Dialogue is that it links up these various, mostly national, initiatives into a more coherent global system. It could play a bridging role across levels of governance, to help reconcile international and regional level frameworks with national approaches and ensure perspectives from actual implementation, particularly in inherently complex sectors like health, inform global standards._ The AI Dialogue can become a platform of integrated and synergetic AI governance moving beyond parallel actions to real cooperation and true global alignment.

It is important to motivate a variety of stakeholders to engage in the AI Dialogue on the basis of their careers and local responsibilities.Resulting governance frameworks need to be relevant, and policymakers are therefore indispensable to coordinate regulatory approaches. Researchers can offer empirical data, evaluation standards and tool reviews.Civil society has a crucial role to play in focusing on the ethical, inclusive and social implications of AI.In fields such as healthcare, bringing in doctors is particularly essential.Clinical decisions are becoming heavily reliant on AI, but if researchers open up the black-box to data-driven solutions and methodologies without the influence of those who work with patients directly, it could lead to a risk of developing solutions that are unimplementable, or simply not relevant in practice. Format-wise, the Dialogue need not become philosophical. Plenary sessions can help bring alignment on shared priorities, but this is an area that even with the best of intentions does not lead to meaningful advancement; this is going to come from smaller task-oriented groups by topic. These working groups should be both multi-disciplinary and oriented around tangible outputs, rather than points of discussion. Perhaps, even more valuable would be the discussion of real world cases. Knowing how AI works—or does not work—in the real world can open our eyes to insurmountable gaps that are not always obvious at the theoretical level. Lastly, the Dialogue should not be a one-time event. We need some kind of followup — working groups that keep digging into the details and/or checkpoints every 6 months or so to review the ideas and status of implementation. The Dialogue would be of a more practical and continuous nature, making it more relevant and powerful.

There are many important perspectives not fully accounted for in global AI governance discussions. Front-line practitioners play the least role in this field, particularly in healthcare. To include their voices in governance requires aimed consultations or working groups by sector, so that policy is more anchored in reality.Second, low- and middle-income countries are underrepresented. Many of the governing frameworks are shaped by high-income settings, which may not reflect the different resource constraints, infrastructure or population needs that others face. For voices to truly represent it is not only a requisite to be invited but that the organisational support has to appear too in form of funding, accessible formats and long term engagement.Patient and end-user perspectives Another gap is the patient s perspective. AI has a major impact on medicine: It directly affects patient care, but the views of patients themselves rarely appear in discussions about risk, trust and acceptability. Structured involvement by patients representatives and advocacy groups would make governance more complete and solid in the eyes of the public.Furthermore, early-career professionals and people whose work is at the intersection of different disciplines—such as health, data science, and policy—are often overlooked. These groups are actively involved in implementation and at the cutting edge of innovation, so they have valuable lessons for emerging challenges.To incorporate these perspectives, AI Dialogues need to go beyond traditional representation models and use methods that are more flexible, such as open calls, the spike (i.e. open participation), or creating dedicated spaces for specific communities, including an agenda that explicitly caters to them.

As a result, the AI Dialogue should take on more interactive, and practice- orientated forms than simply traditional panels or round tables.On such case-based sessions, for example, real-world examples are presented and discussed by multidisciplinary teams. For example, in the medical field studying how an AI tool actually functions in a clinical setting--both success stories and blind alleys--is likely to produce more down-and mid-level insights than purely theoretical discussions could offer.Another potentially useful format is mini-workshops led by specialists on a particular topic with specific goals such as drafting recommendations or discovering holes in major areas. Such forums can help palpitation of ideas and inputs from a wide range more so than trying to do things all-in-one large plenary discussion.Situation-based exercises or actually putting different policies into practice is another idea. By studying scenarios in reality--say a high-quality AI system with uncertain performance is rolled out as just one instance to explore how decisions get made, responsibility can be allocated and risks managed.Participants involved in each sector must also create forums for dialogue that cross vertical areas.both within the fields in themselves as well as between intersecting fields Face-to face exchanges help to break down silos and provide more balanced discussions. Having ongoing digital platforms or communities as part of the Dialogue helps to foster continuous exchange between meetings. This would allow members to share experiences, follow up on ideas, and build longer-term collaboration over time.Mainly, those formats that are interactive, multi-disciplinary and rooted in real-world practice are likely to achieve practical results.

Several existing methods now offer realistic insight into implementing AI governance effectively. The European Union's approach to regulation based on risks can be seen as a prescriptive example. By classifying AI systems in terms of risk and by such an overall orientation, it offers a structure that balances innovation with safety - particularly important for high-impact applications like healthcare. In parallel, organizations like the Organization for Economic Co-Operation and Development and United Nations Educational, Scientific and Cultural Organization hand-principles sets of rules out hope on basic standards for healthy AI (which have been widely accepted and together form a cross-country ethical foundation). Their strength is that they provide common language and direction, across very different regulatory environments. From a more practical perspective, regulatory sandboxes are an important tool. These allow AI systems to be tested in regulated real-world settings and this in turn helps shift accountability earlier, while also encouraging innovation. This type of approach is especially valuable in fields like healthcare which require validation within the real world. Another fruitful area of work is the creation of privacy-preserving data-sharing models such as federated learning. Practices like this allow institutions to cooperate together without having to hand over actual data, thereby responding both to data protection questions and the fact that there is a need for larger sets of more representative data. In healthcare specifically, emerging attempts to set standards for clinical validation and reporting of AI models are crucial. This shift with respect to retrospective performance metrics, more towards robust real-world evaluation. In sum, it is not a single solution but a combination of legal frameworks, technical approaches and practices that develop with use; which shows effective AI governance.